def test_config_init(self, train_config):
c = train_config
config = PPO.generate_config({})
config["frame_config"]["models"] = ["Actor", "Critic"]
config["frame_config"]["model_kwargs"] = [
{
"state_dim": c.observe_dim,
"action_num": c.action_num
},
{
"state_dim": c.observe_dim
},
]
ppo = PPO.init_from_config(config)
old_state = state = t.zeros([1, c.observe_dim], dtype=t.float32)
action = t.zeros([1, 1], dtype=t.float32)
ppo.store_episode([{
"state": {
"state": old_state
},
"action": {
"action": action
},
"next_state": {
"state": state
},
"reward": 0,
"terminal": False,
} for _ in range(3)])
ppo.update()
def gail(self, train_config, device, dtype):
# not used for training, only used for testing apis
c = train_config
actor = smw(
Actor(c.observe_dim, c.action_num).type(dtype).to(device), device,
device)
critic = smw(
Critic(c.observe_dim).type(dtype).to(device), device, device)
discriminator = smw(
Discriminator(c.observe_dim, c.action_num).type(dtype).to(device),
device,
device,
)
ppo = PPO(
actor,
critic,
t.optim.Adam,
nn.MSELoss(reduction="sum"),
replay_device="cpu",
replay_size=c.replay_size,
)
gail = GAIL(
discriminator,
ppo,
t.optim.Adam,
expert_replay_device="cpu",
expert_replay_size=c.replay_size,
)
return gail
def ppo_train(self, train_config):
c = train_config
actor = smw(Actor(c.observe_dim, c.action_num), "cpu", "cpu")
critic = smw(Critic(c.observe_dim), "cpu", "cpu")
ppo = PPO(actor,
critic,
t.optim.Adam,
nn.MSELoss(reduction='sum'),
replay_device="cpu",
replay_size=c.replay_size)
return ppo
def ppo(self, train_config, device, dtype):
c = train_config
actor = smw(
Actor(c.observe_dim, c.action_num).type(dtype).to(device), device,
device)
critic = smw(
Critic(c.observe_dim).type(dtype).to(device), device, device)
ppo = PPO(actor,
critic,
t.optim.Adam,
nn.MSELoss(reduction='sum'),
replay_device="cpu",
replay_size=c.replay_size)
return ppo
def ppo_vis(self, train_config, tmpdir):
# not used for training, only used for testing apis
c = train_config
tmp_dir = tmpdir.make_numbered_dir()
actor = smw(Actor(c.observe_dim, c.action_num)
.to(c.device), c.device, c.device)
critic = smw(Critic(c.observe_dim)
.to(c.device), c.device, c.device)
ppo = PPO(actor, critic,
t.optim.Adam,
nn.MSELoss(reduction='sum'),
replay_device=c.device,
replay_size=c.replay_size,
visualize=True,
visualize_dir=str(tmp_dir))
return ppo
def gail_lr(self, train_config, device, dtype):
# not used for training, only used for testing apis
c = train_config
actor = smw(
Actor(c.observe_dim, c.action_num).type(dtype).to(device), device,
device)
critic = smw(
Critic(c.observe_dim).type(dtype).to(device), device, device)
discriminator = smw(
Discriminator(c.observe_dim, c.action_num).type(dtype).to(device),
device,
device,
)
ppo = PPO(
actor,
critic,
t.optim.Adam,
nn.MSELoss(reduction="sum"),
replay_device="cpu",
replay_size=c.replay_size,
)
lr_func = gen_learning_rate_func([(0, 1e-3), (200000, 3e-4)],
logger=logger)
with pytest.raises(TypeError, match="missing .+ positional argument"):
_ = GAIL(
discriminator,
ppo,
t.optim.Adam,
expert_replay_device="cpu",
expert_replay_size=c.replay_size,
lr_scheduler=LambdaLR,
)
gail = GAIL(
discriminator,
ppo,
t.optim.Adam,
expert_replay_device="cpu",
expert_replay_size=c.replay_size,
lr_scheduler=LambdaLR,
lr_scheduler_args=((lr_func, ), ),
)
return gail
def gail_train(self, train_config):
c = train_config
actor = smw(Actor(c.observe_dim, c.action_num), "cpu", "cpu")
critic = smw(Critic(c.observe_dim), "cpu", "cpu")
discriminator = smw(Discriminator(c.observe_dim, c.action_num), "cpu",
"cpu")
ppo = PPO(
actor,
critic,
t.optim.Adam,
nn.MSELoss(reduction="sum"),
replay_device="cpu",
replay_size=c.replay_size,
)
gail = GAIL(
discriminator,
ppo,
t.optim.Adam,
expert_replay_device="cpu",
expert_replay_size=c.replay_size,
)
return gail